Advancing Embodied Intelligence in Robotic-Assisted Endovascular Procedures: A Systematic Review of AI Solutions

TL;DR

This systematic review examines how Embodied Intelligence can transform robotic-assisted endovascular procedures by unifying intelligent perception and learning-based control within a perception-decision-control loop. It identifies key perception components (vascular segmentation, device tracking, and multi-modal registration) and learning-based control paradigms (RL, imitation, and hybrids), while highlighting data scarcity, lack of benchmarks, and the translation gap as major obstacles. The authors propose a roadmap prioritizing augmented intelligence, development of foundation models and standardized benchmarks, patient-specific adaptation, and robust regulatory/ethical frameworks to enable safe clinical deployment. The work emphasizes a shift from full autonomy toward clinician-supervised, intelligent tooling that enhances precision, safety, and accessibility of endovascular care.

Abstract

Endovascular procedures have revolutionized vascular disease treatment, yet their manual execution is challenged by the demands for high precision, operator fatigue, and radiation exposure. Robotic systems have emerged as transformative solutions to mitigate these inherent limitations. A pivotal moment has arrived, where a confluence of pressing clinical needs and breakthroughs in AI creates an opportunity for a paradigm shift toward Embodied Intelligence (EI), enabling robots to navigate complex vascular networks and adapt to dynamic physiological conditions. Data-driven approaches, leveraging advanced computer vision, medical image analysis, and machine learning, drive this evolution by enabling real-time vessel segmentation, device tracking, and anatomical landmark detection. Reinforcement learning and imitation learning further enhance navigation strategies and replicate expert techniques. This review systematically analyzes the integration of EI into endovascular robotics, identifying profound systemic challenges such as the heterogeneity in validation standards and the gap between human mimicry and machine-native capabilities. Based on this analysis, a conceptual roadmap is proposed that reframes the ultimate objective away from systems that supplant clinical decision-making. This vision of augmented intelligence, where the clinician's role evolves into that of a high-level supervisor, provides a principled foundation for the future of the field.

Figures (6)

• Figure 1: Overview of endovascular intervention, robotic assistance, and the proposed EI framework. (a) Endovascular procedure: flexible instruments (guidewires and catheters) are introduced percutaneously via radial or femoral access and navigated within the vascular lumen under X‑ray fluoroscopy to reach targets such as coronary stenoses or aneurysms for diagnosis and therapy. (b) Robotic-assisted workflow: a leader–follower architecture is depicted in which the operator at a cockpit (leader console) issues push–pull, rotate, and torque commands that are executed by a bedside robotic drive (follower) manipulating the instruments. A C‑arm X‑ray system provides real-time imaging, and a vision monitor displays imaging and system status. The operating table supports the patient; optional sensors (e.g., encoders, force/torque, haptics) enable safety monitoring and feedback. (c) Conceptual framework for EI in robotic-assisted endovascular procedures: perception modules (vascular anatomy segmentation, instrument monitoring, and multimodal registration with preoperative imaging) form the system state; a decision layer uses learning-based policy optimization (e.g., reinforcement learning) with task models and safety constraints to select high-level actions; a low-level controller executes device motions subject to hardware limits. Training and evaluation occur in both simulation and the real operating room; a patient-specific digital twin can model anatomy, device-vessel interaction, imaging geometry, policy learning, and online adaptation. Arrows indicate information and control flow. (d) Composition of EI: core capabilities include vascular segmentation and topology extraction, device detection, tracking, and tip localization, multimodal 2D/3D registration for spatial context, and learning-based control for navigation tasks such as branch cannulation, lesion crossing, and device delivery. Key technical challenges include real-time integration, generalization across vendors and anatomies, robustness to motion and low contrast, and uncertainty estimation; ethical and clinical considerations include patient safety, radiation management, data privacy, and regulatory compliance, with expert-in-the-loop supervision for safe autonomy. The schematic was created with BioRender (https://biorender.com).
• Figure 2: Chord diagram depicting the interplay between perception and control modules in robotic-assisted endovascular procedures. Perception modules (blue) include vessel segmentation, instrument detection and tracking, and multi-modal registration, linked to control modules (purple) via clinical integration. Control encompasses autonomous navigation. Arrows denote directional relationships, with line thickness reflecting connection strength, illustrating their integrated roles in enhancing procedural outcomes.
• Figure 3: PRISMA flowchart illustrating the study selection process. A total of 138 articles were screened for eligibility. Of these, 26 were excluded due to irrelevance to the study topic, insufficient methodological quality, or incomplete data. Ultimately, 112 studies were included in the final review.
• Figure 4: Perception tasks in endovascular procedures. (a) Intraoperative vessel segmentation from DSA images to obtain detailed vascular anatomical structures iyer2021angionet. (b) Device detection and localization utilizing intraoperative DSA images, providing essential visual feedback for physicians and robotic systems zhou2020lightweight. (c) Multi-modal image registration of preoperative Computed tomography angiography (CTA) and intraoperative DSA for enhanced navigation citesong2024iterative. (d) Vessel segmentation from intraoperative DSA images using a GAN-based architecture with contrastive learning ma2021self. (e) Endpoint-based CFKD-Net with multi-branch feature aggregation for weakly-supervised segmentation ma2023towards. (f) Diffusion-based method for progressive vessel structure refinement kim2024c. (g) Fast recurrent attention network for simultaneous mono- and dual-guidewire segmentation and tracking zhou2020lightweight. (h) Coordinate regression-based deep learning for automated catheter detection aghasizade2023coordinate. Images reproduced from iyer2021angionetzhou2020lightweightsong2024iterativema2021selfkim2024cma2023towards and aghasizade2023coordinate.
• Figure 5: Learning frameworks for robotic-assisted endovascular procedures can be categorized as follows: (a) The traditional reinforcement learning (RL) singh2022reinforcement framework for endovascular navigation involves a robotic agent learning through direct interaction with the vascular environment. This approach utilizes target networks and experience buffers to optimize catheter/guidewire manipulation policies robertshaw2023artificialkarstensen2022learningsong2022learningtian2023ddpgscarponi2024zero. (b) The demonstration-based learning framework incorporates expert cardiologist demonstrations to guide the learning of robotic manipulation policies. This method enables more efficient acquisition of complex endovascular navigation skills li2023casogzhou2022learningli2024model. (c) The human-in-the-loop learning framework integrates expert cardiologist supervision into the robotic training process. This approach combines automated learning with clinical expertise to enhance safety and efficacy in endovascular procedures jianu2024cathsimjianu2024autonomous.